Big Day For North American AI Policy: A U.S. Bill, A Sprawling Canadian Plan
June 5, 2026
D.A.D. today covers 15 stories. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
The Daily AI Digest is a daily AI briefing automated by Alexander Panetta — a veteran political journalist tracking the field during a Master's in AI Management at Georgetown University.
D.A.D. Joke of the Day: I asked Claude to help me cut my presentation in half. It removed the second half and called it "efficient summarization."
What's New
AI developments from the last 24 hours
Bipartisan Bill Would Make AI Giants Guard Against Bioweapons and Rogue Models—and Bar States From Their Own Rules
The biggest AI labs would have to test their most powerful models for "catastrophic risk"—a foreseeable path to 50-plus deaths or $1 billion in damage, whether by helping build a chemical or biological weapon, powering an autonomous cyberattack, or slipping free of human control—then publish what they find, mitigate the dangers, and open their books to twice-yearly audits by licensed outside inspectors. Fall short, and the fines run up to $1 million a day. That's the core of the Great American Artificial Intelligence Act, a bipartisan House discussion draft from Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.)—the most serious bipartisan attempt yet to put real guardrails on the AI frontier.
But that safety regime is welded to a bargain igniting a left-right backlash: in return, Washington would bar states from regulating how AI models are built—a three-year preemption that still lets them govern how AI is used. As POLITICO reports, the clause is drawing fire from all sides: AI-safety groups say it turns the floor on state AI rules into a ceiling, red-state Republicans like Ron DeSantis and Rep. Byron Donalds have blasted federal preemption, and Massachusetts and New York Democrats are pressing Trahan to abandon it. The 269-page draft also creates a Commerce Department AI-standards office, protects AI whistleblowers, stiffens penalties for AI-enabled fraud, and funds workforce training.
Why it matters: This is the grand bargain at the heart of U.S. AI policy: mandatory safety oversight of the largest labs in exchange for one national rulebook instead of fifty. POLITICO calls it "the last realistic chance to craft federal rules" before the midterms—but it's still just a discussion draft, and the preemption clause meant to win Republicans is the very thing uniting safety hawks and states'-rights conservatives against it. Who gets to govern AI—and whether Washington can agree before the models grow more dangerous—may turn on this fight.
Source: obernolte.house.gov · POLITICO
Canada Helped Build Modern AI. Its New Strategy Admits It Fell Behind on Using It.
Canada helped invent modern AI—researchers like Geoffrey Hinton, Yoshua Bengio and Richard Sutton are among the field's founding figures—yet it has fallen badly behind at actually using the technology: just 12% of Canadian businesses had adopted AI by 2024–25 (versus 29–42% across the Nordics), and while Canada ranks 5th in the world for AI venture capital, roughly 70% of its AI startups decamp abroad. Prime Minister Mark Carney's new national strategy, "AI for All," is the catch-up plan, and it's unusually candid about the gap—aiming to lift business adoption from 12% to 60% by 2034.
The substance is a striking blend of demand-side help and old-fashioned industrial policy. A new AI Missions Program opens with $200 million to improve health outcomes, paired with a $100 million Health Sector Data Space (with the Canadian Institute for Health Information) to put public-system health data to work; more missions are promised in priority sectors like energy and agriculture. Ottawa would also pour $700 million into subsidized compute for small firms, build a "world-leading" public supercomputer by 2031, and lead a global effort to invest in and sustain open-source AI in the public interest. Most striking is the interventionist turn: a $500 million Canadian Tech Growth Fund that would let the federal government "at times" take direct equity stakes in the most promising Canadian AI firms. The strategy also stakes out Canada's natural niche—trusted AI for highly regulated, data-sensitive industries like finance, health care and telecom, where sovereignty and security matter most—with Toronto's Cohere as the enterprise-and-government champion and Bengio's LawZero flying the safety-by-design flag.
The harder problem is trust. Canada ranks 42nd of 47 countries on public confidence in AI, so a plan built on mass adoption has to win over a deeply skeptical public—which the strategy tries to do with free AI-literacy training for every Canadian and a promised tightening of privacy and consumer law: online-safety rules, a "fundamental right to privacy," and curbs on AI-driven "surveillance pricing."
Why it matters: For the many Canadians in this audience, it's the country's defining AI bet of the decade—and it's already drawing fire from both sides. The left calls it too business-focused, a productivity agenda wearing a public-interest coat. The right says it ducks the real reasons capital leaves Canada—tangled energy permitting for data centers and a tax regime unfriendly to capital investment—while leaning on alliances (the new Canada–Germany Sovereign Technology Alliance) to paper over compute, chip and cloud gaps it can't close alone. The wager underneath it all: that Canada's problem was never the science—it was the will to adopt.
Source: ised-isde.canada.ca · PM launch announcement
Anthropic: Our AI Is Starting To Generate Itself, And Here Are The Implications
AI is starting to build AI. In a report titled "When AI builds itself," Anthropic says it has crossed a threshold researchers have anticipated for decades: its own models now do the bulk of the work to create the next generation of models—the first turn of a flywheel known as recursive self-improvement. The numbers it discloses are vertiginous. More than 80% of the code Anthropic now ships is written by Claude, up from low single digits in early 2025. The length of a task a model can finish on its own is doubling every four months—from four-minute jobs to twelve-hour ones inside a year. On an internal research test, Claude went from a 3x speedup to a 52x one in twelve months, work the authors flatly call "superhuman." In one trial, AI agents cracked 97% of an open AI-safety problem that two human researchers had managed only 23% of in a week.
The most haunting parts are human. Anthropic's own engineers, quoted anonymously, describe a creeping loss of purpose—one hasn't hand-written code in five months; another confesses that "on days where everything works well, I can't help but think nothing I do matters." If the loop fully closes, with AI autonomously designing its own successor, the company concedes the odds of humans losing control rise sharply—and argues the world should preserve the option to pause frontier development before that point. This is a frontier lab stating, on the record, that the technology it sells may be slipping beyond human authorship.
So read it with both eyes open. Anthropic itself calls recursive self-improvement "not inevitable," and nearly every jaw-dropping figure comes from its own unpublished internal data: lines-of-code counts it admits overstate the gains, success rates graded by a Claude "judge," and a staff survey whose own cited research finds engineers routinely overestimate how much AI helps. It also lands as Anthropic moves toward a reported IPO and days before its next flagship, Mythos—whose preview build is the hero of nearly every chart. A company raising money on the story that AI is about to remake the world is not a neutral witness to it.
Why it matters: If even the cautious version of this is true, the timeline to transformative AI is compressing—reshaping the economy, the labor market, and the regulatory fights now underway faster than institutions can react. But the report is also a case study in how the biggest AI claims now reach us: self-reported, hard to verify, and inseparable from the commercial interests of the lab making them. The job is to hold both at once—take the trajectory seriously without taking Anthropic's framing at face value.
Discuss on Hacker News · Source: anthropic.com
The Open-Source Defection the AI Labs Fear: Lindy Dumps Anthropic for DeepSeek
Yesterday we flagged the open-source defection risk hanging over the labs' coming IPOs; today it got a vivid data point. Flo Crivello, CEO of the fast-growing AI-agent startup Lindy, posted that he "pulled the trigger" and switched 100% of Lindy's traffic to the open-weight Chinese model DeepSeek v4, "churning from Anthropic models." The part that should worry Anthropic is his claim that it's not just cheaper but better: the move "saves us millions of $ and we're actually seeing an increase in performance on many core use cases. Transformative for the business." Investor Brandon Carl noted Lindy is backed by Menlo, Coatue and Battery, and predicted the lesson will spread to those firms' other portfolio companies.
It's one more enterprise voice—after Uber's Dara Khosrowshahi mused last week about moving "to open source"—telling the same story: once an open-weight model is good enough, the premium for a closed frontier model gets hard to justify. Here the customer says the open option already wins on both price and performance.
Why it matters: This is exactly why Anthropic has so much riding on its next flagship, Mythos, blowing people away. The closed labs' trillion-dollar IPO theses assume enterprises will keep paying a premium for proprietary models for years—but every Lindy that defects to open weights, and claims better results for less, chips at that premise. If "good enough and far cheaper" becomes the default before the labs reach profitability, they may never recoup the tens of billions they're spending to stay at the frontier.
Source: X (@Altimor) · @brandonjcarl
Cloudflare Acquires Vite, a Cornerstone of Modern Web Development
Cloudflare is acquiring VoidZero, the company behind Vite—a widely-used build tool for web development that powers many JavaScript and TypeScript projects. VoidZero founder Evan You and the full team will join Cloudflare. The company says Vite and related projects (Vitest, Rolldown, Oxc) will remain open source, MIT-licensed, and vendor-agnostic. Cloudflare is committing $1 million to a Vite ecosystem fund for maintainers and contributors.
Why it matters: For teams building with Vite, the immediate promise is continuity—but the acquisition signals Cloudflare's ambition to own more of the developer toolchain, which could shape how web apps are built and where they're deployed.
Discuss on Hacker News · Source: blog.cloudflare.com
What's in the Lab
New announcements from major AI labs
ChatGPT Gets Smarter Memory That Synthesizes Your History Over Time
OpenAI is overhauling how ChatGPT remembers you. The update is the most capable version yet of an approach it calls "dreaming," which synthesizes memories in the background from your full chat history rather than only saving the discrete facts you explicitly flag. OpenAI built it around three goals: carrying useful context forward, following stated preferences and constraints, and—the hardest part—staying current as time passes, so "you're going to Singapore in July" updates to "you went to Singapore in July 2026" once the trip ends instead of leaving the model stuck on stale context. A new, reviewable memory summary page lets you see the highlights of what ChatGPT has inferred about you, correct or delete entries, and tell it which topics to raise and when. Plus and Pro users in the US get it today; a roughly 5x cut in the compute needed to run dreaming clears the way to extend it to Free and Go users—and to raise memory limits for paying users—over the coming weeks.
Why it matters: OpenAI is betting that persistent, self-updating memory—not just conversation-by-conversation context—is what makes an assistant genuinely useful over years, and a real switching cost against rivals. The editable memory summary is also its answer to the obvious worry: a system quietly inferring a profile of you from everything you type is only palatable if you can see and correct what it thinks it knows.
OpenAI Unveils Biodefense Strategy, Betting on 'Responsible Defenders'
OpenAI released a biodefense action plan outlining its strategy for deploying AI in biological security. The plan accompanies two previously announced tools: GPT-Rosalind (April 2026), a reasoning model for biology research, and Rosalind Biodefense (May 2026), designed to help vetted developers build pandemic preparedness applications. OpenAI argues that equipping "responsible defenders" with advanced AI—rather than restricting access broadly—is the best approach to biosecurity. The company says it's developing safeguards and governance frameworks, though specifics weren't detailed.
Why it matters: This signals OpenAI's bet that controlled access beats restriction in dual-use domains—a position that will likely draw scrutiny as AI capabilities in biology advance and regulators weigh how to handle potential misuse risks.
Meta Engineers Data Centers to Survive Instant Power Failures
Meta has built what it calls "Instantaneous PowerLoss Storm" testing into its data center infrastructure—a system designed to handle sudden, zero-warning power failures. The company says resilience to instant outages is now engineered across its entire stack, from physical facilities to server racks to its Twine container orchestrator. The approach includes battery-backed data persistence and an internal alert system called Power Loss Siren. No performance benchmarks were disclosed.
Why it matters: As AI workloads grow more complex and expensive to interrupt, Meta is signaling that catastrophic failure scenarios are now a first-class engineering concern—infrastructure decisions that could shape reliability expectations across the industry.
Codex Expands Beyond Developers With Plugins for Analysts, Marketers, Designers
OpenAI is repositioning Codex beyond coding. The tool now has 5 million weekly users, with non-developers making up 20% of that base—and that segment is growing three times faster than developers. New features include six role-specific plugins covering analysts, marketers, designers, researchers, investors, and bankers, bundling 62 apps and 110 skills. OpenAI also previewed the ability to create shareable interactive websites and apps via URL. Annotations let users refine outputs iteratively.
Why it matters: OpenAI is betting that AI coding tools have broader appeal than 'coding'—this signals a push to make Codex a general-purpose work assistant competing directly with enterprise workflow tools.
Anthropic Open-Sources AI Security Scanner, But Labels It a Research Artifact
Anthropic has released a framework on GitHub that enables AI agents to hunt for security vulnerabilities in code. The tool can run up to 10 parallel agents scanning codebases, though the repository is marked as unmaintained and not accepting contributions—suggesting this is more research artifact than production tool. Community reaction has been skeptical: developers estimate running costs in the hundreds to thousands of dollars depending on the model used, and some question why Anthropic would open-source this rather than monetize it directly if it worked well.
Why it matters: The release signals major AI labs are exploring automated security auditing, but the unmaintained status and high costs suggest this capability isn't ready to replace traditional vulnerability scanners—yet.
Discuss on Hacker News · Source: github.com
What's in Academe
New papers on AI and its effects from researchers
Most AI Risks Carry 'Intolerable' Odds of Catastrophe, Say 272 of the Field's Experts
Ask the people who build and study AI to put numbers on the danger, and the picture is grim. In a new MIT FutureTech–University of Queensland study, 272 AI experts from 37 countries scored 24 categories of AI risk—and judged that, on the current trajectory, 18 of them carry a greater than 10% chance of catastrophe within five years, where "catastrophe" means more than a million deaths or $100 billion in damage. "If we consider other mature areas of technology, such as nuclear power or aviation, risks at that level would be treated as intolerable," says Neil Thompson, who directs MIT FutureTech. Even assuming sensible, cost-effective safeguards get adopted, five risks stay above that 10% line—led by dangerous capabilities, AI-enabled weapons and cyberattacks, and power centralization.
The experts flagged a second, structural problem: the people most exposed to AI's harms—ordinary users and the public—have the least power to prevent them, while the actors with that power (frontier developers, governments, regulators) carry the responsibility. That mismatch, co-author Peter Slattery argues, is why "laws, treaties, and other collective-action mechanisms" may be needed where voluntary self-governance falls short. One honest caveat from the authors: these are expert judgments, not measured odds, and the per-risk probabilities overlap—they shouldn't be summed into a single doomsday number.
Why it matters: A 10% expert-estimated chance of catastrophe isn't a forecast—but it's a sobering signal from the people who understand these systems best, and it lands the same week as Anthropic's recursive-self-improvement paper and the Great American AI Act's "catastrophic risk" provisions. Three very different documents, one converging worry: that competitive pressure is the core danger, and coordination the hardest part of any fix.
Source: mitsloan.mit.edu · Data: osf.io
AI Hedge Funds' Early Edge Has Disappeared, Major Study Finds
A new working paper distributed by NBER examined AI-driven investing using regulatory filings and fund disclosures, finding the strategy has grown steadily since the early 2010s and concentrates in hedge funds. The surprising finding: AI hedge funds initially outperformed their conventional peers, but this edge has eroded over time—even among early adopters. One counterintuitive result challenges a common worry: AI funds actually showed less similar returns to each other than traditional funds, suggesting the technology isn't creating herding behavior.
Why it matters: For executives evaluating AI-driven funds, the research suggests the 'alpha' from machine learning strategies may be temporary as adoption spreads—a pattern familiar from other quantitative investing waves.
Economic Model Shows Why AI Competition May Push Firms Past Safety Limits
In a new working paper distributed by NBER, economists lay out a theoretical model analyzing how market competition shapes AI existential risk. The paper argues that AGI safety depends heavily on market structure—not just technical factors. Its central finding: above a certain market size threshold, firms will race to develop AGI even when doing so has negative expected value for society. The model treats safety spending as a resource allocation problem, where competitive pressure systematically pushes firms toward speed over caution.
Why it matters: This is academic ammunition for policymakers considering whether the AI industry's structure—not just its technology—requires regulatory intervention.
Complex Economic Modeling That Took Days Now Runs in Minutes
Researchers Victor Duarte and Julia Fonseca have developed an AI method for solving complex economic equilibrium models—the kind used to estimate how markets respond to policy changes or shocks. Their approach uses neural networks trained on simulated data and claims to compress estimation times from days to under 20 minutes. In one test, the traditional method couldn't match the AI's accuracy even after four days of computation. The system also includes an AI agent that can apply the method to new economic models from plain-language prompts.
Why it matters: For economists, finance teams, and policy analysts who run structural models, this signals that AI may soon eliminate one of the field's major bottlenecks—the computational cost that limits how many scenarios you can test.
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Five Years Later, 'Stochastic Parrots' Warnings Have Largely Come True
A retrospective argues that warnings in the 2020 paper that led to Timnit Gebru's firing from Google have been validated by subsequent events. 'On the Dangers of Stochastic Parrots,' co-authored with Emily Bender, flagged risks including hallucinations, bias amplification, environmental costs, and unauditable training data. The article cites Google's emissions up 48% since 2019, Microsoft's up 29%, and the discovery that LAION-5B (used to train Stable Diffusion) contained thousands of CSAM images. Gebru's departure became a flashpoint in debates over AI ethics research independence.
Why it matters: The piece reflects ongoing tension between AI labs' commercial incentives and internal safety research—a dynamic that continues to shape how companies handle dissent and which concerns get aired publicly before products ship.
Discuss on Hacker News · Source: tumblr.com
What's Happening on Capitol Hill
Upcoming AI-related committee hearings
Thursday, June 11 — Hearings to examine AI and the American dream, focusing on promoting innovation, affordability and American dominance. Senate · Senate Banking, Housing, and Urban Affairs (Open Hearing) 538, Dirksen Senate Office Building
What's On The Pod
Some new podcast episodes
AI in Business — How Industrial Service Leaders Are Closing the Knowledge Gap Before It's Too Late with Mike Hughes of Peak International Group
The Cognitive Revolution — Nested Learning: Ali Behrouz on the Quest for Continual Learning & Illusion of AI Architectures
How I AI — Gemini Omni: Clone yourself with AI in under 15 minutes